Add the first working version of the preconditioned CGM.
authorMichael Orlitzky <michael@orlitzky.com>
Wed, 20 Mar 2013 03:45:05 +0000 (23:45 -0400)
committerMichael Orlitzky <michael@orlitzky.com>
Wed, 20 Mar 2013 03:45:05 +0000 (23:45 -0400)
optimization/preconditioned_conjugate_gradient_method.m [new file with mode: 0644]
tests/preconditioned_conjugate_gradient_method_tests.m [new file with mode: 0644]

diff --git a/optimization/preconditioned_conjugate_gradient_method.m b/optimization/preconditioned_conjugate_gradient_method.m
new file mode 100644 (file)
index 0000000..1442194
--- /dev/null
@@ -0,0 +1,96 @@
+function [x, k] = preconditioned_conjugate_gradient_method(A,
+                                                          M,
+                                                          b,
+                                                          x0,
+                                                          tolerance,
+                                                          max_iterations)
+  %
+  % Solve,
+  %
+  %   Ax = b
+  %
+  % or equivalently,
+  %
+  %   min [phi(x) = (1/2)*<Ax,x> + <b,x>]
+  %
+  % using the preconditioned conjugate gradient method (14.56 in
+  % Guler). If ``M`` is the identity matrix, we use the slightly
+  % faster implementation in conjugate_gradient_method.m.
+  %
+  % INPUT:
+  %
+  %   - ``A`` -- The coefficient matrix of the system to solve. Must
+  %     be positive definite.
+  %
+  %   - ``M`` -- The preconditioning matrix. If the actual matrix used
+  %     to precondition ``A`` is called ``C``, i.e. ``C^(-1) * Q *
+  %     C^(-T) == \bar{Q}``, then M=CC^T. However the matrix ``C`` is
+  %     never itself needed. This is explained in Guler, section 14.9.
+  %
+  %   - ``b`` -- The right-hand-side of the system to solve.
+  %
+  %   - ``x0`` -- The starting point for the search.
+  %
+  %   - ``tolerance`` -- How close ``Ax`` has to be to ``b`` (in
+  %     magnitude) before we stop.
+  %
+  %   - ``max_iterations`` -- The maximum number of iterations to
+  %     perform.
+  %
+  % OUTPUT:
+  %
+  %   - ``x`` - The solution to Ax=b.
+  %
+  %   - ``k`` - The ending value of k; that is, the number of
+  %   iterations that were performed.
+  %
+  % NOTES:
+  %
+  % All vectors are assumed to be *column* vectors.
+  %
+  % The cited algorithm contains a typo; in "The Preconditioned
+  % Conjugate-Gradient Method", we are supposed to define
+  % d_{0} = -z_{0}, not -r_{0} as written.
+  %
+  % REFERENCES:
+  %
+  %   1. Guler, Osman. Foundations of Optimization. New York, Springer,
+  %   2010.
+  %
+  n = length(x0);
+
+  if (isequal(M, eye(n)))
+    [x, k] = conjugate_gradient_method(A, b, x0, tolerance, max_iterations);
+    return;
+  end
+
+  zero_vector = zeros(n, 1);
+
+  k = 0;
+  x = x0; % Eschew the 'k' suffix on 'x' for simplicity.
+  rk = A*x - b; % The first residual must be computed the hard way.
+  zk = M \ rk;
+  dk = -zk;
+
+  for k = [ 0 : max_iterations ]
+    if (norm(rk) < tolerance)
+       % Success.
+       return;
+    end
+
+    % Unfortunately, since we don't know the matrix ``C``, it isn't
+    % easy to compute alpha_k with an existing step size function.
+    alpha_k = (rk' * zk)/(dk' * A * dk);
+    x_next = x + alpha_k*dk;
+    r_next = rk + alpha_k*A*dk;
+    z_next = M \ r_next;
+    beta_next = (r_next' * z_next)/(rk' * zk);
+    d_next = -z_next + beta_next*dk;
+
+    k = k + 1;
+    x = x_next;
+    rk = r_next;
+    zk = z_next;
+    dk = d_next;
+  end
+end
diff --git a/tests/preconditioned_conjugate_gradient_method_tests.m b/tests/preconditioned_conjugate_gradient_method_tests.m
new file mode 100644 (file)
index 0000000..c58eb55
--- /dev/null
@@ -0,0 +1,31 @@
+A = [5,1,2; ...
+     1,6,3; ...
+     2,3,7];
+
+M = eye(3);
+
+b = [1;2;3];
+
+x0 = [1;1;1];
+
+## Solved over the rationals.
+cgm  = conjugate_gradient_method(A, b, x0, 1e-6, 1000);
+pcgm = preconditioned_conjugate_gradient_method(A, M, b, x0, 1e-6, 1000);
+diff = norm(cgm - pcgm);
+
+unit_test_equals("PCGM agrees with CGM when M == I", ...
+                true, ...
+                norm(diff) < 1e-6);
+
+
+## Needs to be symmetric!
+M = [0.97466, 0.24345, 0.54850; ...
+     0.24345, 0.73251, 0.76639; ...
+     0.54850, 0.76639, 1.47581];
+
+pcgm = preconditioned_conjugate_gradient_method(A, M, b, x0, 1e-6, 1000);
+diff = norm(cgm - pcgm);
+
+unit_test_equals("PCGM agrees with CGM when M != I", ...
+                true, ...
+                norm(diff) < 1e-6);